【摘要】 专利分类是专利分析的基本任务,而基于深度学习的自动化专利分类方法可以有效地执行专利分类任务。现有研究大多利用自然语言处理方法,基于单个专利的文本内容(如摘要和标题)对专利文本进行分类,而专利及标签之间的宏观关系(如专利之间的引用和标签之间的共现)在很大程度上被忽略。为了缓解专利分类中单一专利信息的问题,构建了三个图网络表示专利及其标签之间的宏观关系,然后提出一个基于混合表征学习的专利分类框架,将专利及标签的宏观关系融入分类中,以提高专利的自动化分类的准确性。在真实的专利数据集的实验结果表明,该分类方法在多个评价指标上取得了最佳的性能。更多还原

【Abstract】 Patent classification is a fundamental task in patent analysis, and automatic patent classification methods based on deep learning can perform effectively. Most existing studies utilize natural language processing methods to classify patent documents based on the textual contents of individual patents( e. g., abstracts and titles), while the macro-relationships between patents and their labels( e. g., citations between patents and co-occurrence between labels) are largely ignored. To alleviate the problem of single patent information in patent classification, this paper constructed three graphs to represent the macro-relationships between patents and labels, and then proposed a hybrid representation based framework for patent classification to enhance the automated classification of patents. Extensive experiments conducted on real patent datasets show that the proposed classification method achieves the best performance in several evaluation metrics.更多还原